STeMP: Spatio-Temporal Modelling Protocol

Spatio-temporal predictive modelling is a key method in the geosciences. Often, machine-learning, which can be applied to complex, non-linear and interacting relationships, is preferred over classical (geo)statistical models. However, machine-learning models are often perceived as "black boxes", meaning that it is hard to understand their inner workings. Furthermore, there are several pitfalls associated with the application of machine-learning models in general, and spatio-temporal machine-learning models in particular. This might, e.g., concern the spatial autocorrelation inherent in spatial data, which complicates data splitting for model validation. Following from this, it is key to transparently report spatio-temporal models. Transparent reporting can facilitate interpreting, evaluating and reproducing spatio-temporal models, and can be used to determine their suitability for a specific research question. Standardized model protocols are particularly valuable in this context, as they document model parameters, decisions and assumptions. While such protocols exist for machine-learning models in general (e.g., Model Cards, REFORMs), as well as for specific domains like species distribution modelling (ODMAP), such protocols are lacking in the general field of spatio-temporal modelling. Here, we present ideas for STeMP (Spatio-Temporal Modelling Protocol), a protocol for spatio-temporal models that fills this gap. The protocol is designed to be beneficial for all parties involved in the modeling process, including model developers, maintainers, reviewers, and end-users. The protocol is implemented as a web application and is structured in three sections: Overview, Model and Prediction. The Overview section contains general metadata, while the following two sections go into more detail. The Model section includes modules describing, for example, the predictors, model validation procedures, and software. The optional Prediction section contains information about the prediction domain, map evaluation, and uncertainty assessment.To make the protocol useful during model development, warnings are raised when common pitfalls are encountered (e.g., if an unsuitable cross-validation strategy is used). These warnings can be automatically retrieved from a filled protocol, spotlighting potential issues and helping authors and reviewers. Moreover, we provide the optional possibility to generate automated reports and also inspection figures from user-provided inputs (e.g., from model objects as well as from training and test data sets). The protocol is hosted on GitHub (https://github.com/LOEK-RS/STeMP) and hence open to flexible incorporation of feedback from the broader community.With our presentation, we aim to encourage the discussion of our proposed model report in the spatio-temporal modelling community.